Frontier models leave a fingerprint, and it's the lab's, not the model's (www.pedronasc.com)

A recent study has demonstrated that the writing style of frontier AI models, specifically OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 and 4.8, is distinctively influenced by the lab that developed them rather than inherent to the models themselves. A classifier trained on 25 stylistic features accurately distinguished GPT from the Claude models with 87-93% accuracy using a small set of prompts. Key stylistic differences include Claude's frequent use of em-dashes and contractions, compared to GPT's heavier reliance on semicolons and commas. Notably, the classifier couldn’t differentiate between the two Claude versions, suggesting a consistent internal style within Anthropic despite their behavioral updates. This finding is significant for the AI/ML community as it highlights the importance of understanding the "house style" of AI models in addition to their capabilities. As benchmark tests primarily measure a model's performance on specific tasks, this study encourages a reevaluation of how we assess model outputs. By recognizing that two models with similar capabilities can have fundamentally different defaults in phrasing and refusal strategies, users can better choose AI systems based on desired behaviors rather than just performance scores. This work sets the stage for further research into the evolving idiosyncrasies of AI writing styles across different providers.